- 學位論文
利用類神經網路預測心臟冠狀動脈鈣化
Using Artificial Neural Network to Predict Coronary Artery Calcification
摘要
冠狀動脈心血管疾病總稱冠心病(Coronary Artery Disease，CAD)，常見的原因是因血管壁內膜與外膜間的夾層中堆積粥狀斑塊，使血管壁斑塊的積聚增加(血管壁再造)形成鈣化(calcification)，造成冠狀動脈血流減少，進而容易引起心肌梗塞。 本研究從2008年11月至西元2013年12月間共蒐集328例資料，以多列偵測電腦斷層掃描(Multidetector Computed Tomography，MDCT)驗證利用類神經網路(Artificial Neural Network，ANN)對於三種冠狀動脈鈣化(coronary artery calcification)事件狀況進行了分析預測，研擬了四種預測模組並與臨床上常用的佛來明罕危險分數(Framingham Risk Score；FRS)風險評分系統進行效能評估並比較之，每一預測模組皆以研究資料隨機分為80%訓練組與20%測試組。 其結果與預測效能如下：(1)在預測冠狀動脈是否有鈣化的風險上，類神經網路預測模組一與模組二之接受者操作特徵曲线下面積(Area under Receiver Operating Characteristic Curve，AUROC)分別為0.91與0.83，預測效能皆優於佛來明罕危險分數的0.771，由此可知類神經網路對於冠狀動脈是否鈣化預測，優於佛來明罕危險分數；(2)預測冠狀動脈鈣化積分(Coronary Artery Calcium Score，CACS)中，類神經網路預測模組三之相關係數0.793為高度相關預測模組，其預測效能優於冠狀動脈鈣化積分與佛來明罕危險分數之相關係數0.291；(3)最後在預測冠狀動脈鈣化積分為基礎的6項風險分級上，類神經網路預測模組四與佛來明罕危險分數雖說無相同的效能比較，但由於類神經網路測模組四中冠狀動脈鈣化樣本數不足故預測效果不好，整體的預測準確率僅有57.63%，而佛來明罕危險分數與冠狀動脈鈣化6項風險分級程度之間的關聯強度(strength of association：ω2)指數達20.29%，二者屬於具強度關係。 總結上述3點，類神經網路預測除了第3點冠狀動脈鈣化風險分級程度因樣本數不足故預測效果不好外，類神經網路預測模組與佛來明罕危險分數相較下有較好預測能力，因此本研究建議以類神經網路模組分析冠狀動脈鈣化事件與發生可能性，並適合臨床使用達到早期有效預防冠心病。
並列摘要
Generally Coronary Artery Disease (CAD) are indicated coronary artery and cardiovascular disease.The most common cause of cardiovascular disease is coronary artery due to the accumulation of the vessel wall in atherosclerotic plaques, laminated between the intima and adventitia, let the blood vessel wall plaque accumulate increased (vessel wall reconstruction) and trigger the formation of calcification , resulted in diminished blood flow of coronary artery, which usually leads to myocardial infarction. The dataset of this study had collected total of 328 cases since November 2008 to December 2013. And this study was applied Multi-detector Computed Tomography (MDCT) for predicting analysis of three subtype coronary artery calcification by Artificial Neural Network (ANN). Developed four predictive models in cooperation with Framingham Risk Score (FRS) to assess and compare the performance. The dataset was randomly subdivided into 80% of training set and 20% of testing set. The result and predictive performance as following:(1) Predict module I and predict module II, in prediction the risk of coronary artery calcification, of the Artificial Neural Network in term of Area under Receiver Operating Characteristic Curve (AUROC) were 0.91 and 0.83, which performance prediction were better than Framingham Risk Score of 0.771, Hence, to predict coronary artery calcification Artificial Neural Network is superior to the Framingham Risk Score.(2) As Coronary Artery Calcium Score (CACS) for prediction, Artificial Neural Network prediction model III with correlation coefficient of 0.793, are highly correlated prediction module, its prediction performance is superior to Coronary Artery Calcification Score and Framingham risk score the correlation coefficient of 0.291;(3) Finally, six risks classification based on Coronary Artery Calcification Score , even though Artificial Neural Network prediction model IV and Framingham Risk Score were not the same in term of comparison performance , but due to the Artificial Neural Network prediction module IV sample size insufficient in category of coronary artery calcification therefore predicted to be ineffective, the overall prediction accuracy rate is only 57.63%, while the strength of association (ω2) index between the Framingham risk score and Six risk category of coronary artery calcification is 20.29 %, both are with positive strength of association. To sum up the above three items, beside item of the third with poor prediction value in coronary artery calcification risk score due to insufficient sample size, the ability of prediction of Artificial Neural Network is better than Framingham Risk Score. Thus, this study suggests that to apply prediction model of Artificial Neural Network to analyze the probability of occurrence of coronary artery calcification events, and is suitable for clinical use to achieve early and effective prevention Coronary Artery Disease (CAD).